Best pose estimation model

[ghost]

Hi @choyingw,
I am trying to use your pose estimation model - the one reproduces the results in the paper (https://drive.google.com/file/d/13LagnHnPvBjWoQwkR3p7egYC6_MVtmG0/view?usp=sharing) - but get fixed pose predicted angles for images with different poses.
When I am using with the regular model you published (https://drive.google.com/file/d/1BVHbiLTfX6iTeJcNbh-jgHjWDoemfrzG/view?usp=sharing) this phenomenon doesn't happen and I get more logical results (but not SOTA for pose estimation).

I was wondering if this happens to you as well and is there a problem in the model you published?

[ghost]

@choyingw
I also took a look into the state_dict of the "best_pose.pth.tar" and found keys with "IGM" prefix (instead of "I2P") - when we cleaned this inconsistency and compared the rest of state dict with the current published PyTorch model class we found keys that are not appeared in what you published. For example the following keys:
'.classifier_pitch.1.bias', '.classifier_pitch.1.weight', '.classifier_roll.1.bias', '.classifier_roll.1.weight', '.classifier_scale.1.bias', '.classifier_scale.1.weight', '.classifier_texture.1.bias', '.classifier_texture.1.weight', '.classifier_trans.1.bias', '.classifier_trans.1.weight', '.classifier_yaw.1.bias', '.classifier_yaw.1.weight',

We believe that the model state dict you published for pose estimation doesn't match the PyTorch model class.

Can you please clarify what to do?

[choyingw]

Oops. I uploaded the wrong model. I've updated the readme and the link. Please check.

[ghost]

@choyingw Thanks for your quick response. I still get different results from what you reported on aflw2000 ( Yaw: 5.537 Pitch: 8.978 Roll: 6.132 || MAE: 6.882).

My implementation is similar to https://github.com/vitoralbiero/img2pose/blob/main/evaluation/jupyter_notebooks/aflw_2000_3d_evaluation.ipynb
load image and pose (from mat file) -> face detection -> bbox+margin (as you did) -> crop -> extract pose

I tried to follow your evaluation code, but found several files that you didn't share. For example, './aflw2000_data/AFLW2000-3D_crop', './aflw2000_data/AFLW2000-3D_crop.list', './aflw2000_data/eval/ALFW2000-3D_pose_3ANG_excl.npy', './aflw2000_data/eval/ALFW2000-3D_pose_3ANG_skip.npy'.

can you please share how you created 'aflw2000_data' folder and share the folder itself using google drive, or alternatively create full evaluation code that use the original data to estimate pose - it will be great.

[choyingw]

aflw2000-3D is shared in the link (ReadME, Single Image Inference Demo - Step4, Download the data)

python benchmark.py -w "pathToPoseModel", you'll get the reported number.

[bigdelys]

@choyingw Even using the latest "best_pose.pth.tar" I am still getting constant poses for any input image, equal to
[ 0.35037747 -4.45007563 -0.32709743]

[choyingw]

@bigdelys Hi, I didn't find this issue on my end. As I print out pose using AFLW2000-3D, the head pose angles are different.